A smart optimization method and system for aluminum alloy die casting process

Abstract

This invention discloses an intelligent optimization method and system for the forming process of aluminum alloy die casting, belonging to the field of intelligent optimization technology. The method includes: real-time acquisition of the multi-element content in the molten aluminum alloy; combining the multi-element content into a three-dimensional material gene vector; inputting the three-dimensional material gene vector into a pre-trained graph neural network to calculate the injection speed compensation coefficient and the boosting pressure compensation coefficient; spatiotemporally aligning the geometric center coordinates of defects with the mold temperature field and pressure time-series data recorded during the die casting process, inputting this data into a reinforcement learning surrogate model constructed using a deep Q-network algorithm to reconstruct the evolution path of the internal defect region; and reversing and updating the weight parameters of the graph neural network based on the defect formation time and location points identified in the evolution path. This invention achieves millisecond-level collaborative compensation of injection speed and boosting pressure by fusing multi-element content into a three-dimensional material gene vector and utilizing the dynamic topology modeling capability of a pre-trained graph neural network.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology, and in particular to an intelligent optimization method and system for the forming process of aluminum alloy die casting parts. Background Technology

[0002] Significant progress has been made in the intelligent control of aluminum alloy die casting processes in recent years. Spectroscopic analysis has enabled millisecond-level online detection of multi-element content in the melt; graph neural networks are maturing in industrial process modeling, capable of handling high-dimensional process parameter mapping problems; and the combination of industrial CT and deep learning can achieve sub-millimeter-level identification of internal defects in die castings. The optimization capabilities of reinforcement learning algorithms in dynamic system control have also been validated through multiple industrial cases, providing a technological foundation for closed-loop process control.

[0003] Existing technologies have shortcomings, lacking real-time modeling capabilities for the dynamic coupling effect between melt composition fluctuations and die-casting parameters. Traditional solutions rely on static process windows or offline calibration models, failing to capture the nonlinear influence of multi-element interactions in molten aluminum alloys on flow / solidification behavior. When raw material batches fluctuate, fixed-parameter strategies lead to a sharp increase in air entrapment and shrinkage cavity defect rates, becoming a bottleneck for high-precision die-casting quality control. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent optimization method for the forming process of aluminum alloy die casting to solve the problem of process parameter mismatch caused by real-time fluctuations in the content of multiple elements in the melt during aluminum alloy die casting.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent optimization method for the forming process of aluminum alloy die casting, which includes real-time acquisition of the content of multiple elements in the aluminum alloy melt and combining the content of multiple elements into a three-dimensional material gene vector.

[0008] The three-dimensional material gene vector is input into a pre-trained graph neural network to calculate the injection velocity compensation coefficient and the boost pressure compensation coefficient.

[0009] Based on the injection speed compensation coefficient and the boost pressure compensation coefficient, the real-time injection speed and boost pressure settings of the die casting machine are adjusted, and the die casting operation is performed to generate the die casting.

[0010] Three-dimensional tomographic images of die-cast parts are obtained by scanning. Internal defect regions in the three-dimensional tomographic images are identified by deep learning segmentation algorithms, and four-dimensional defect center coordinates are generated.

[0011] The coordinates of the geometric center of the defect are spatiotemporally aligned with the mold temperature field and pressure time series data recorded during the die casting process. The input is used to construct a reinforcement learning surrogate model through a deep Q-network algorithm to reconstruct the evolution path of the internal defect region.

[0012] Based on the defect formation time and location points identified in the evolution path, the weight parameters of the graph neural network are corrected and updated in reverse.

[0013] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting forming process described in this invention, the specific steps for combining the contents of multiple elements into a three-dimensional material gene vector are as follows:

[0014] A steady melt flow is formed by driving the melt with an axial magnetic field and breaking the surface oxide film. The steady melt flow is then discretized into a uniform melt droplet array under the shearing of an inert gas.

[0015] A ring plasma is excited by a pulsed laser with vortex phase modulation applied to a uniform molten droplet array, and the spectral signal generated by the plasma is collected to calculate the content of multiple elements.

[0016] Input the multi-element content into the aluminum-based alloy phase diagram database, call the phase diagram parameters, construct the element interaction matrix based on atomic properties and extract the main eigenvalues ​​of element interactions;

[0017] The content of multiple elements, phase diagram parameters, and element interaction principal feature values ​​are combined to form a three-dimensional material gene vector.

[0018] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting forming process described in this invention, the specific steps for calculating the injection speed compensation coefficient and the boosting pressure compensation coefficient are as follows:

[0019] The three-dimensional material gene vector is input into a pre-trained graph neural network to generate a dynamic physical constraint graph structure. The vector dimension is mapped to graph nodes, and edge weights are constructed based on multi-scale physical coupling relationships.

[0020] A quantum projection-driven graph convolution operation is applied to a dynamic physical constraint graph structure, and global features are generated by aggregating them through complex feature space transformation.

[0021] The global features are decoded using both thermodynamic and fluid dynamic methods to output the injection velocity compensation coefficient and the boost pressure compensation coefficient, respectively.

[0022] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting process described in this invention, the specific steps for performing the die-casting operation to generate the die-cast part are as follows:

[0023] Based on the injection velocity compensation coefficient and the boost pressure compensation coefficient, calculate the dynamic gain coefficient of injection velocity and the dynamic gain coefficient of boost pressure.

[0024] The dynamic gain coefficients of injection speed and boost pressure are combined with the basic process parameters to solve the linkage constraint and generate the adjusted set values ​​of injection speed and boost pressure.

[0025] The adjusted injection speed and boost pressure settings are synchronously sent to the servo valve group via real-time bus to execute the die-casting operation and generate the die-cast part.

[0026] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting forming process described in this invention, the specific steps for generating the four-dimensional defect center coordinates are as follows:

[0027] Dual-energy tomography was performed on the die-cast parts to obtain three-dimensional tomographic images, and the three-dimensional tomographic images were then processed by tomographic reconstruction to generate three-dimensional volume data.

[0028] Three-dimensional volume data is processed using deep learning segmentation algorithms to identify internal defect areas;

[0029] Extract the geometric center coordinates of the internal defect region and integrate them with the die-casting filling time to generate four-dimensional defect center coordinates.

[0030] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting forming process described in this invention, the specific steps for reconstructing the evolution path of the internal defect region are as follows:

[0031] In the mold coordinate system, the four-dimensional defect center coordinates are integrated with the mold temperature field and pressure time series data recorded during the die casting process, and spatiotemporal alignment is performed to construct five-dimensional spatiotemporal field data;

[0032] Five-dimensional spatiotemporal field data is input and a reinforcement learning agent model is constructed through a deep Q-network algorithm. Evolutionary action sequences are then explored and generated through physical constraint strategies.

[0033] The evolutionary action sequence is regularized and inverted using multi-field coupled partial differential equations to reconstruct the evolutionary path of the internal defect region.

[0034] As a preferred embodiment of the intelligent optimization method for the aluminum alloy die-casting forming process described in this invention, the weight parameters of the reverse-corrected graph neural network are updated, and the specific steps are as follows:

[0035] Based on the evolution path of the reconstructed internal defect region, a holographic defect energy field is constructed in the five-dimensional anti-de Sitter spacetime, with the defect formation time point mapped as time coordinates and the location point mapped as boundary space coordinates;

[0036] The holographic defect energy field is input into a non-commutative differential operator, and the spatiotemporal invariants are extracted as quantum geometric features through quantum covariant gradient;

[0037] A pre-trained string theory matrix model is used to perform dimensionality renormalization on quantum geometric features to generate a modified tensor for the weights of the graph neural network.

[0038] The modified tensor is injected into the backpropagation flow of the graph neural network to perform the update.

[0039] Secondly, the present invention provides an intelligent optimization system for the forming process of aluminum alloy die casting, including an element acquisition module, an intelligent compensation module, a linkage execution module, a defect analysis module, a time-tracing inversion module, and a reverse update module;

[0040] The element acquisition module is used to collect the content of multiple elements in the molten aluminum alloy in real time and combine the content of multiple elements into a three-dimensional material gene vector.

[0041] The intelligent compensation module is used to input the three-dimensional material gene vector into a pre-trained graph neural network to calculate the injection speed compensation coefficient and the pressurization pressure compensation coefficient.

[0042] The linkage execution module is used to adjust the real-time injection speed and boosting pressure setting values ​​of the die casting machine based on the injection speed compensation coefficient and the boosting pressure compensation coefficient, and to perform the die casting operation to generate the die casting.

[0043] The defect analysis module is used to scan the die-cast casting to obtain a three-dimensional tomographic image, identify the internal defect region in the three-dimensional tomographic image through a deep learning segmentation algorithm, and generate four-dimensional defect center coordinates.

[0044] The time-tracing inversion module is used to align the geometric center coordinates of the defect with the time-series data of mold temperature field and pressure recorded during the die-casting process in time and space. The input is used to construct a reinforcement learning surrogate model through a deep Q-network algorithm to reconstruct the evolution path of the internal defect region.

[0045] The reverse update module is used to reverse-correct the weight parameters of the graph neural network and update them based on the defect formation time and location points identified in the evolution path.

[0046] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent optimization method for the aluminum alloy die-casting forming process as described in the first aspect of the present invention.

[0047] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent optimization method for the aluminum alloy die-casting forming process as described in the first aspect of the present invention.

[0048] The beneficial effects of this invention are as follows: It integrates the content of multiple elements into a three-dimensional material gene vector, and through the dynamic topology modeling capability of a pre-trained graph neural network, achieves millisecond-level collaborative compensation between injection speed and pressurization pressure. Utilizing the multi-scale physical coupling mechanism of graph convolution, the accuracy-speed compensation coefficient and pressure compensation amount directly drive the real-time control of the die-casting machine. Simultaneously, it constructs a physically interpretable mapping relationship between the material gene and process parameters, providing real-time optical diagnostic-level decision support for intelligent die casting. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart of an intelligent optimization method for the forming process of aluminum alloy die casting parts.

[0051] Figure 2 A module diagram of an intelligent optimization system for aluminum alloy die casting forming process.

[0052] Figure 3 This is a flowchart of the graph neural network compensation calculation.

[0053] Figure 4 A flowchart for reconstructing the evolutionary path. Detailed Implementation

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0057] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an intelligent optimization method for the forming process of aluminum alloy die casting, comprising the following steps:

[0058] S1. Real-time acquisition of the multi-element content in molten aluminum alloy, and combination of the multi-element content into a three-dimensional material gene vector.

[0059] Furthermore, the axial magnetic field drives the melt to form a steady melt flow and breaks the surface oxide film, discretizing the steady melt flow into a uniform melt droplet array under the shearing of inert gas;

[0060] Specifically, a vertical Lorentz force is applied to the conductive melt by an axial magnetic field, inducing the formation of a self-organized rotating flow field inside the melt. Under the Coriolis effect, the self-organized rotating flow field is stretched along the axis and develops into a steady-state vortex structure. The radial velocity gradient generated by the steady-state vortex structure exceeds the oxide film binding energy threshold, realizing the oxide layer rupture and peeling. The ruptured melt is sheared by the inert gas injected by the coaxial annular nozzle. The inert gas shear flow tears the continuous melt into a uniform array of melt droplets.

[0061] It should be noted that the oxide film bonding energy threshold is set based on the experimental determination of the bonding strength of the cleavage plane of aluminum oxide crystals on the aluminum alloy surface, and the example value is 2.4 joules per square meter.

[0062] A ring plasma is excited by a pulsed laser with vortex phase modulation applied to a uniform molten droplet array, and the spectral signal generated by the plasma is collected to calculate the content of multiple elements.

[0063] Specifically, a vortex phase-modulated pulsed laser is focused on the zero point of the rotating phase of the molten droplet array. The diameter of the pulsed laser spot matches the spacing of the molten droplet array, increasing the energy density of the pulsed laser. The molten droplet array absorbs the pulsed laser energy and undergoes multiphoton ionization. Multiphoton ionization causes electron avalanche breakdown at the outer edge of the molten droplet array, generating a ring plasma. The ring plasma expands and cools, and the bound state electrons are stimulated to emit atomic characteristic spectral lines. A high-resolution fiber optic spectrometer simultaneously acquires the emission spectrum data of the ring plasma during the cooling stage. The emission spectrum data is compared with the standard spectral database of aluminum alloy materials to calculate the percentage content of multiple elements such as silicon, iron, copper, manganese, and zinc.

[0064] Input the multi-element content into the aluminum-based alloy phase diagram database, call the phase diagram parameters, construct the element interaction matrix based on atomic properties and extract the main eigenvalues ​​of element interactions;

[0065] Specifically, the aluminum-based alloy phase diagram database receives the percentage content of multiple elements such as silicon, iron, copper, manganese, and zinc. The aluminum-based alloy phase diagram database indexes the binary eutectic point, solidification shrinkage rate, dendrite spacing, and solubility parameters that match the percentage content of multiple elements. The atomic property database calls the electronegativity atomic radius and valence electron number data of silicon, iron, copper, manganese, and zinc. According to Pauling's rule, the orbital hybridization energy and band center offset and lattice distortion energy between each pair of elements are calculated to form an element interaction matrix. The maximum modulus eigenvalue of the element interaction matrix is ​​extracted as the principal eigenvalue of the element interaction.

[0066] It should be noted that the aluminum-based alloy phase diagram database refers to the database that stores binary phases of alloy systems composed of aluminum, silicon, iron, copper, manganese, and zinc. Figure 3 A metallurgical database containing phase diagrams, multi-phase diagrams, phase equilibrium points during solidification, solidification shrinkage, dendrite spacing, and solubility parameters; an atomic property database, referring to a quantum chemical database storing atomic numbers, electron configurations, orbital hybridization, electronegativity, atomic radii, number of valence electrons, and first ionization energy of silicon, iron, copper, manganese, zinc, and aluminum.

[0067] The content of multiple elements, phase diagram parameters, and element interaction principal feature values ​​are combined to form a three-dimensional material gene vector.

[0068] Specifically, the multi-element content is transferred to the first dimension of the three-dimensional material gene vector, the phase diagram parameters called by the aluminum-based alloy phase diagram database are embedded in the second dimension of the three-dimensional material gene vector, the element interaction principal feature value is located to the last dimension of the three-dimensional material gene vector, the format of the three-dimensional material gene vector conforms to the fixed-length floating-point array specification, and the generation of the three-dimensional material gene vector completes the spatial combination mapping of multi-element content, phase diagram parameters and element interaction principal feature values.

[0069] S2. Input the three-dimensional material gene vector into the pre-trained graph neural network to calculate the injection velocity compensation coefficient and the boost pressure compensation coefficient.

[0070] Furthermore, the three-dimensional material gene vector is input into a pre-trained graph neural network to generate a dynamic physical constraint graph structure, which maps the vector dimension to graph nodes and constructs edge weights based on multi-scale physical coupling relationships.

[0071] It should be noted that the three-dimensional material gene vector is used as the node feature to initialize the graph neural network. The edge weights of the graph neural network are assigned initial values ​​according to the physical topology connection rules of melt flow. The graph convolution operation aggregates the thermal conductivity coefficient and rheological stress state of adjacent nodes layer by layer. The fully connected layer at the output end maps the refined features to the dual target space composed of the injection velocity compensation coefficient and the pressurization pressure compensation coefficient. During the training process, the loss function composed of the velocity compensation coefficient residual, the pressure compensation coefficient residual and the defect inversion consistency constraint is minimized. Backpropagation is performed to perform stochastic gradient descent optimization with Nesterov momentum. Finally, a graph neural network with high-precision prediction capability of injection velocity compensation coefficient and pressurization pressure compensation coefficient is obtained.

[0072] Specifically, a dynamic physical constraint graph structure is generated, which maps vector dimensions to graph nodes and constructs edge weights based on multi-scale physical coupling relationships. The expression is as follows:

[0073] ;

[0074] In the formula, Indicates the source node and target node edge weights, Indicates the source node index. Indicates the target node index. This refers to the Softplus activation function. Indicates a spatial scale hierarchical index. This indicates that the summation is performed across all spatial scales. Indicating spatial scale classification The scale weight coefficient, Indicating spatial scale classification Spatial correlation function, This represents the Euclidean distance between nodes. Represents the gain coefficient of the integral term. Indicates the interval of scale factor Accumulate points. This represents the maximum value of the scaling factor. This represents the minimum value of the scaling factor. Represents the Gaussian kernel function. Indicating spatial scale classification The decay function, Indicates the scale factor. The differential variable representing the scaling factor.

[0075] It should be noted that spatial scale classification The scale weight coefficients are obtained through backpropagation training of the graph neural network. Example values ​​are... , , The integral term gain coefficient was obtained by Bayesian optimization from multiphysics coupling matching experimental data, with an example value range of 0.25-0.35.

[0076] A quantum projection-driven graph convolution operation is applied to a dynamic physical constraint graph structure, and global features are generated by aggregating them through complex feature space transformation.

[0077] Specifically, the quantum projection operator acts on the node-embedded complex vectors of the dynamic physical constraint graph structure, and projects the node complex vectors onto the entangled Hilbert space through unitary matrix transformation. The linear combination features of the entangled Hilbert space are phase-modulated by Pauli gated rotation. The phase-modulated entangled features are subjected to dot product operation by the complex domain graph convolution kernel. The complex domain graph convolution kernel is matched with the adjacency tensor of the dynamic physical constraint graph structure to complete local feature aggregation. The local aggregated features are filtered by Fourier frequency domain to extract long-range correlation components. The long-range correlation components are restored to the coordinate space by inverse quantum Fourier transform. The restored features are projected onto the real subspace for dimensionality reduction to generate global features.

[0078] Specifically, the global features are decoded using both thermodynamic and fluid dynamic methods, outputting the injection velocity compensation coefficient and the boost pressure compensation coefficient, respectively, with the following expressions:

[0079] ;

[0080] ;

[0081] In the formula, This represents the injection velocity compensation coefficient. The reference value representing the injection velocity compensation coefficient. The scaling factor represents the injection velocity compensation factor. This represents the sigmoid function. Represented as a fully connected neural network for speed decision-making. This is represented as a global eigenfunction renormalization function. Represents the global feature vector. This represents the boost pressure compensation coefficient. This represents the baseline value for the boost pressure compensation coefficient. The scaling factor represents the boost pressure compensation coefficient. Represents the hyperbolic tangent function. Represented as a fully connected neural network for stress-based decision-making. This represents the normalization function.

[0082] It should be noted that the injection speed compensation coefficient is determined through industrial-grade die-casting test calibration and feedback optimization, with an example value range of 0.82 to 1.18; the baseline value of the injection speed compensation coefficient is based on the fluid dynamics calculation of the material's critical filling speed, with an example fixed baseline value of 0.8; the scaling factor of the injection speed compensation coefficient is obtained by fitting the wall thickness-flow rate nonlinear equation, with an example fixed value of 0.4; the boosting pressure compensation coefficient is optimized based on the clamping force-internal pressure coupling model, with an example value range of 0.55 to 1.42; the baseline value of the boosting pressure compensation coefficient corresponds to the material's solidification shrinkage pressure value, with an example fixed baseline value of 1.0; and the scaling factor of the boosting pressure compensation coefficient is set through statistical regression analysis of shrinkage defects, with an example fixed value of 0.5.

[0083] S3. Based on the injection speed compensation coefficient and the boost pressure compensation coefficient, adjust the real-time injection speed and boost pressure settings of the die casting machine, and execute the die casting operation to generate the die-cast casting.

[0084] Specifically, based on the injection velocity compensation coefficient and the boost pressure compensation coefficient, the dynamic gain coefficients of injection velocity and boost pressure are calculated, and their expressions are as follows:

[0085] ;

[0086] ;

[0087] In the formula, This represents the dynamic gain coefficient of the injection velocity. Indicates the thermal sensitivity coefficient. Indicates the temperature difference of hydraulic oil. This indicates the nonlinear effect of adjusting the injection velocity compensation coefficient. This represents the dynamic gain coefficient of boost pressure. This represents the absolute value of the pressure deviation. Indicates deviation, Indicates pressure, Represents the natural logarithm function. This indicates the nonlinear effect of adjusting the boost pressure compensation coefficient.

[0088] It should be noted that the speed dynamic gain coefficient is based on the experimental calibration of hydraulic oil temperature-viscosity characteristics, and the example is optimized to 1.6 to 2.3 in the oil temperature range of 50 to 100℃; the heat sensitivity coefficient is obtained by fitting the viscosity-temperature curve Arrhenius equation, and the example is fixed at 0.1; the boost pressure dynamic gain coefficient is optimized by the step response test of the pressure servo valve, and the example is set to 0.42.

[0089] Specifically, the dynamic gain coefficients of injection speed and boost pressure are combined with the basic process parameters for linkage constraint solution to generate the adjusted injection speed and boost pressure setpoints, expressed as follows:

[0090] ;

[0091] In the formula, This represents the minimum value. This indicates the adjusted injection speed setting. Indicates the base injection velocity. Describing the L2 norm operator, This indicates the speed reference after gain adjustment. This is expressed as the squared deviation between the adjusted injection velocity setpoint and the velocity reference after gain adjustment. This represents the optimization weighting coefficient for the stress term. This indicates the adjusted boost pressure setting. Indicates the base boost pressure. This indicates the pressure reference after gain adjustment. This is expressed as the squared deviation between the adjusted boost pressure setpoint and the pressure reference after gain adjustment. Represents the constraint identifier. Indicates the wear coefficient of the servo valve. Indicates the change in velocity. Indicates time interval, Represents the absolute value constraint of the rate of change of velocity. Indicates the pressure gradient. Represents the gradient operator, Indicates the material flow coefficient. This represents the adjusted injection speed setting raised to the power of 0.7.

[0092] It should be noted that the basic process parameters refer to the five core control variables that are preset: the reference injection speed value, the reference boost pressure value, the maximum clamping force threshold, the fast injection stroke length, and the alloy pouring temperature. The maximum clamping force threshold is set based on the safety margin of the product of the casting's projected area and the peak internal pressure, and the example value is 23,000 kN.

[0093] It should be noted that the optimization weight coefficient of the pressure term was determined through bi-objective Pareto front analysis, and the example value is fixed at 0.7; the servo valve wear coefficient was calibrated based on the slope decay rate of the servo valve current-displacement curve, and the example value is 0.35; the material flow coefficient was fitted based on the high shear rate test of the melt rheometer, and the example value is 0.15±0.02.

[0094] The adjusted injection speed and boost pressure settings are synchronously sent to the servo valve group via real-time bus to execute the die-casting operation and generate the die-cast part.

[0095] Specifically, the adjusted injection speed setting and boost pressure setting are encapsulated into EtherCAT real-time bus communication frames. These EtherCAT real-time bus communication frames are synchronously transmitted to the injection speed control servo valve controller and the boost pressure control servo valve controller of the die casting machine at a cycle of less than 500 microseconds. The injection speed control servo valve controller analyzes the setting values ​​and drives the valve core displacement of the injection speed control servo valve. The boost pressure control servo valve controller responds to the setting values ​​and adjusts the opening of the boost pressure control servo valve. The displacement command of the injection speed control servo valve is linked to the plunger advance speed of the die casting machine's injection cylinder. The opening command of the boost pressure control servo valve is linked to the hydraulic pressure output of the boost accumulator. The plunger advance speed of the injection cylinder and the hydraulic pressure of the boost accumulator are precisely matched according to the setting values ​​during the high-pressure injection stroke stage. The molten aluminum alloy completes the mold cavity filling under the linked controlled injection pressure and filling speed. The liquid metal achieves sequential solidification in the final pressure stage maintained by the boost pressure control servo valve. After the metal solidifies, the ejection mechanism separates the mold to generate the die-cast casting.

[0096] S4. Scan the die-cast part to obtain a three-dimensional tomographic image, identify the internal defect region in the three-dimensional tomographic image through a deep learning segmentation algorithm, and generate four-dimensional defect center coordinates.

[0097] Furthermore, dual-energy tomography is performed on the die-cast casting to obtain three-dimensional tomographic images, and the three-dimensional tomographic images are then processed by tomographic reconstruction to generate three-dimensional volume data.

[0098] Specifically, the die-cast part is fixed on the turntable of the dual-energy spectroscopy tomography (DST) equipment. The turntable rotates the die-cast part at a constant speed of 360 degrees. The high-energy X-ray source and the low-energy X-ray source of the DST equipment emit pulses alternately. The high-energy X-ray source pulses penetrate the die-cast part and are received by the high-energy linear array detector, while the low-energy X-ray source pulses penetrate the die-cast part and are received by the low-energy linear array detector. The DST equipment records the projection data of the high-energy linear array detector and the low-energy linear array detector. The projection data is input into the FDK back-projection reconstruction algorithm to perform attenuation coefficient matrix calculation. The attenuation coefficient matrix is ​​corrected for scattering artifacts by Hilbert spatial filtering. After correction, the attenuation coefficient matrix is ​​mapped to three-dimensional space gray voxels. The three-dimensional space gray voxels are arranged in Cartesian coordinate system to generate three-dimensional volume data.

[0099] It should be noted that the attenuation coefficient matrix is ​​a two-dimensional numerical array that is reconstructed from dual-energy spectral tomography projection data and characterizes the ability of materials at various spatial locations inside the casting to attenuate X-rays.

[0100] Three-dimensional volume data is processed using deep learning segmentation algorithms to identify internal defect areas;

[0101] Specifically, the three-dimensional volume data is input into the deep learning segmentation algorithm, which includes a three-dimensional encoder-decoder structure. In the three-dimensional encoder stage, three-dimensional convolutional kernels are used to extract defect features in layers, and the spatial resolution is reconstructed through three-dimensional deconvolution operations. The output layer of the deep learning segmentation algorithm applies the Sigmoid activation function to generate a defect probability map, which is then segmented to form a binary defect mask. A three-dimensional connected component labeling algorithm is executed to identify internal defect regions, and the geometric center coordinates of the internal defect regions are extracted and fused with the die-casting filling time to generate four-dimensional defect center coordinates.

[0102] S5. The coordinates of the geometric center of the defect are spatiotemporally aligned with the mold temperature field and pressure time series data recorded during the die casting process. The input is used to construct a reinforcement learning agent model through a deep Q-network algorithm to reconstruct the evolution path of the internal defect region.

[0103] Furthermore, by integrating the four-dimensional defect center coordinates with the mold temperature field and pressure time series data recorded during the die casting process in the mold coordinate system, spatiotemporal alignment is performed to construct five-dimensional spatiotemporal field data;

[0104] Specifically, a spatial reference point is set at the origin of the mold coordinate system. The spatial components of the four-dimensional defect center coordinates are transformed to the Cartesian coordinates of the mold coordinate system. The mold temperature field time series data recorded during the die casting process is rigidly registered and aligned with the mold coordinate system. The pressure time series data is filled with the time axis through linear interpolation to form a time-continuous pressure field. The pressure field and the mold temperature field are spatiotemporally synchronized by interpolating according to the grid points of the mold coordinate system. The timestamp of the four-dimensional defect center coordinates is associated with the pressure field synchronization time node. The data points of the pressure field synchronization time node are spatially matched with the temperature field of the grid points of the mold coordinate system. The matched defect center coordinates, mold temperature field data points and pressure field data points are combined to form five-dimensional data points. The five-dimensional data points are arranged in lexicographical order to generate five-dimensional spatiotemporal field data.

[0105] Five-dimensional spatiotemporal field data is input and a reinforcement learning agent model is constructed through a deep Q-network algorithm. Evolutionary action sequences are then explored and generated through physical constraint strategies.

[0106] It should be noted that the five-dimensional spatiotemporal field data is input to the deep Q-network algorithm. The deep Q-network algorithm uses a three-dimensional convolutional kernel to process the spatial dimension and a long short-term memory network to process the temporal dimension. The action space is defined as the discrete adjustment value of the injection speed compensation coefficient and the pressurization pressure compensation coefficient. The physical constraint strategy constrains the action selection through the boundary restriction function. The target network adopts a soft update mechanism. The training process performs exploration optimization through a greedy strategy to generate a reinforcement learning agent model that supports the physical constraint strategy.

[0107] Specifically, the reinforcement learning agent model is constructed using a deep Q-network algorithm to input five-dimensional spatiotemporal field data. The spatiotemporal features are extracted by combining the deep Q-network algorithm of the reinforcement learning agent model with a convolutional long short-term memory network. The physical constraint strategy of the deep Q-network algorithm is embedded with action boundary constraint functions, such as truncation functions. The action boundary constraint functions of the physical constraint strategy force the selection of actions to meet the solidification shrinkage rate and pressure gradient constraints of the die casting process. The exploration strategy of the reinforcement learning agent model adopts a greedy method to iteratively select actions and generate an evolutionary action sequence.

[0108] The evolutionary action sequence is regularized and inverted using multi-field coupled partial differential equations to reconstruct the evolutionary path of the internal defect region.

[0109] Specifically, the evolution action sequence is input into a multi-field coupled partial differential equation inversion framework. The multi-field coupled partial differential equations include the Navier-Stokes equations describing melt flow, the Fourier heat conduction equations describing temperature propagation, and the solidification kinetics equations describing phase transition processes. The injection velocity compensation coefficient of the evolution action sequence drives the update of the boundary conditions of the Navier-Stokes equations, and the pressurization pressure compensation coefficient synchronously updates the solidification pressure term of the solidification kinetics equations. The time discretization step size is aligned with the time resolution of the five-dimensional spatiotemporal field data to form an inversion time grid. The regularization constraint uses the Tikhonov norm to penalize non-physical fluctuations in the evolution path. The Newton-Raphson iterative method is used to solve the implicit scheme of the multi-field coupled partial differential equations. In each iteration, the temperature gradient is calculated based on the heat conduction equation to update the position of the solidification front. The spatial overlap between the position of the solidification front and the internal defect region is used as the inversion convergence criterion. The output of the solidification front displacement path after convergence is the evolution path for reconstructing the internal defect region.

[0110] S6. Based on the defect formation time and location points identified in the evolution path, reverse-correct the weight parameters of the graph neural network and update them.

[0111] Furthermore, based on the evolution path of the reconstructed internal defect region, a holographic defect energy field is constructed in five-dimensional anti-de Sitter spacetime, with the defect formation time point mapped as time coordinates and the location point mapped as boundary space coordinates;

[0112] Specifically, the time point data of the evolution path of the reconstructed internal defect region is mapped to the time coordinate dimension of the five-dimensional anti-de Sitter spacetime, and the location point data of the evolution path of the reconstructed internal defect region is projected to the boundary space coordinate dimension of the five-dimensional anti-de Sitter spacetime. The time coordinate dimension and the boundary space coordinate dimension form a discrete spacetime point set in the five-dimensional anti-de Sitter spacetime. Each discrete spacetime point set is associated with the defect energy density value. The negative curvature manifold relationship between the boundary space coordinate points is calculated through the hyperbolic distance function. The negative curvature manifold relationship is converted into a quantum entangled state in the volume interval of the five-dimensional anti-de Sitter spacetime. The quantum entangled state evolves according to the constraints of Einstein's field equations. The evolution process generates a scalar field distribution in the entire five-dimensional anti-de Sitter spacetime, which is expressed as a holographic defect energy field.

[0113] The holographic defect energy field is input into a non-commutative differential operator, and the spatiotemporal invariants are extracted as quantum geometric features through quantum covariant gradient;

[0114] Specifically, the holographic defect energy field is input into the domain of the non-commutative differential operator. The non-commutative differential operator performs discrete-form Weyl algebra operations to generate a non-commutative differential form, which establishes a quantum covariant derivative with the spin connection structure of the Riemannian manifold. The quantum covariant derivative acts on the scalar value of the holographic defect energy field to generate a quantum covariant gradient field. The quantum covariant gradient field is integrally calculated along the orbit of the Abelian gauge group to calculate the curvature form. The curvature form is mapped to the delam cohomology group through the Chern-Wey isomorphism. The upper closed chain of the delam cohomology group extracts the equivariant closed form that satisfies the Karl-Dan-Killing equation. The intrinsic symmetry of the equivariant closed form generates a spacetime invariant cluster, which is output as a quantum geometric feature.

[0115] A pre-trained string theory matrix model is used to perform dimensionality renormalization on quantum geometric features to generate a modified tensor for the weights of the graph neural network.

[0116] It should be noted that the Gaussian unitary ensemble random matrix of the D0-membrane matrix model is initialized, the supersymmetric action functional of the string theory matrix model is imported, the matrix configuration is generated by the hybrid Monte Carlo sampling algorithm, the supersymmetry is verified by solving the eigenspectrum through the Dirac equation, the gradient of the action functional is optimized by the conjugate gradient method, the convergence is determined based on the supersymmetric conserved load deviation, the convergence matrix configuration is stored to complete the pre-training, and the string theory matrix model is obtained.

[0117] Specifically, the quantum geometric features are input into the gauge field operator of the pre-trained string theory matrix model. The string theory matrix model performs operator product expansion through the non-commutation algebra structure of type IIB superstring theory. The operator product expansion generates a gauge-invariant renormalized flow. The dimension regularization parameter adjusts the degree of freedom contraction of the quantum geometric features. The eigenspace after degree of freedom contraction is mapped to the conformal field boundary through AdS / CFT dual mapping. The supersymmetric Yang-Mills field of the conformal field boundary generates renormalized weights. The renormalized weights are decomposed into third-order tensor forms according to the eigenmode. The third-order tensor forms are represented as modified tensors.

[0118] The modified tensor is injected into the backpropagation flow of the graph neural network to perform the update.

[0119] Specifically, the graph neural network backpropagation flow calculates the weight gradient tensor of the current training step. The graph neural network weight correction tensor matches the dimension of the weight gradient tensor through a dimension alignment operation. The matched graph neural network weight correction tensor completely overwrites the original weight gradient tensor. The overwritten weight gradient tensor is submitted to the stochastic gradient descent optimizer with Nesterov momentum. The stochastic gradient descent optimizer with Nesterov momentum performs parameter updates according to the learning rate scaling factor and momentum accumulation. The parameter update results are synchronized to the graph neural network to complete the update.

[0120] This embodiment also provides an intelligent optimization system for the forming process of aluminum alloy die casting, including: an element acquisition module, an intelligent compensation module, a linkage execution module, a defect analysis module, a time-tracing inversion module, and a reverse update module; the element acquisition module is used to collect the multi-element content in the molten aluminum alloy in real time and combine the multi-element content into a three-dimensional material gene vector; the intelligent compensation module is used to input the three-dimensional material gene vector into a pre-trained graph neural network to calculate the injection speed compensation coefficient and the boost pressure compensation coefficient; the linkage execution module is used to adjust the real-time injection speed and boost pressure setpoints of the die casting machine based on the injection speed compensation coefficient and the boost pressure compensation coefficient. The system performs die casting to generate die-cast parts; the defect analysis module scans the die-cast parts to obtain three-dimensional tomographic images, identifies internal defect regions in the three-dimensional tomographic images using a deep learning segmentation algorithm, and generates four-dimensional defect center coordinates; the time-tracking inversion module aligns the geometric center coordinates of the defects with the time-series data of mold temperature field and pressure recorded during the die casting process, inputs them to construct a reinforcement learning surrogate model using a deep Q-network algorithm, and reconstructs the evolution path of the internal defect region; the reverse update module corrects the weight parameters of the graph neural network and updates it based on the defect formation time and location points identified in the evolution path.

[0121] This embodiment also provides a computer device applicable to the intelligent optimization method for aluminum alloy die casting forming process, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent optimization method for aluminum alloy die casting forming process proposed in the above embodiment.

[0122] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0123] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent optimization method for the aluminum alloy die-casting forming process as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0124] In summary, this invention achieves millisecond-level coordinated compensation of injection speed and pressurization pressure by fusing the content of multiple elements into a three-dimensional material gene vector and leveraging the dynamic topology modeling capability of a pre-trained graph neural network. Utilizing the multi-scale physical coupling mechanism of graph convolution, the accuracy-speed compensation coefficient and pressure compensation amount directly drive the real-time control of the die-casting machine. Simultaneously, a physically interpretable mapping relationship between the material gene and process parameters is constructed, providing real-time optical diagnostic-level decision support for intelligent die casting.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent optimization method for the forming process of aluminum alloy die casting parts, characterized in that: include, Real-time acquisition of multi-element content in molten aluminum alloy, and combination of multi-element content into a three-dimensional material gene vector; The three-dimensional material gene vector is input into a pre-trained graph neural network to calculate the injection velocity compensation coefficient and the boost pressure compensation coefficient. Based on the injection speed compensation coefficient and the boost pressure compensation coefficient, the real-time injection speed and boost pressure settings of the die casting machine are adjusted, and the die casting operation is performed to generate the die casting. Three-dimensional tomographic images of die-cast parts are obtained by scanning. Internal defect regions in the three-dimensional tomographic images are identified by deep learning segmentation algorithms, and four-dimensional defect center coordinates are generated. The coordinates of the four-dimensional defect center are spatiotemporally aligned with the mold temperature field and pressure time series data recorded during the die casting process. The input is used to construct a reinforcement learning surrogate model through a deep Q-network algorithm to reconstruct the evolution path of the internal defect region. Based on the defect formation time and location points identified in the evolution path, the weight parameters of the graph neural network are corrected and updated in reverse.

2. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 1, characterized in that: The specific steps for combining the contents of multiple elements into a three-dimensional material gene vector are as follows. A steady melt flow is formed by driving the melt with an axial magnetic field and breaking the surface oxide film. The steady melt flow is then discretized into a uniform melt droplet array under the shearing of an inert gas. A ring plasma is excited by a pulsed laser with vortex phase modulation applied to a uniform molten droplet array, and the spectral signal generated by the plasma is collected to calculate the content of multiple elements. Input the multi-element content into the aluminum-based alloy phase diagram database, call the phase diagram parameters, construct the element interaction matrix based on atomic properties and extract the main eigenvalues ​​of element interactions; The content of multiple elements, phase diagram parameters, and element interaction principal feature values ​​are combined to form a three-dimensional material gene vector.

3. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 2, characterized in that: The calculations for the injection velocity compensation coefficient and the boost pressure compensation coefficient are performed using the following steps: The three-dimensional material gene vector is input into a pre-trained graph neural network to generate a dynamic physical constraint graph structure. The vector dimension is mapped to graph nodes, and edge weights are constructed based on multi-scale physical coupling relationships. A quantum projection-driven graph convolution operation is applied to a dynamic physical constraint graph structure, and global features are generated by aggregating them through complex feature space transformation. The global features are decoded using both thermodynamic and fluid dynamic methods to output the injection velocity compensation coefficient and the boost pressure compensation coefficient, respectively.

4. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 3, characterized in that: The specific steps for performing the die-casting operation to generate the die-cast part are as follows. Based on the injection velocity compensation coefficient and the boost pressure compensation coefficient, calculate the dynamic gain coefficient of injection velocity and the dynamic gain coefficient of boost pressure. The dynamic gain coefficients of injection speed and boost pressure are combined with the basic process parameters to solve the linkage constraint and generate the adjusted set values ​​of injection speed and boost pressure. The adjusted injection speed and boost pressure settings are synchronously sent to the servo valve group via real-time bus to execute the die-casting operation and generate the die-cast part.

5. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 4, characterized in that: The specific steps for generating the four-dimensional defect center coordinates are as follows: Dual-energy tomography was performed on the die-cast parts to obtain three-dimensional tomographic images, and the three-dimensional tomographic images were then processed by tomographic reconstruction to generate three-dimensional volume data. Three-dimensional volume data is processed using deep learning segmentation algorithms to identify internal defect areas; Extract the geometric center coordinates of the internal defect region and integrate them with the die-casting filling time to generate four-dimensional defect center coordinates.

6. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 5, characterized in that: The evolution path of the reconstructed internal defect region involves the following specific steps. In the mold coordinate system, the four-dimensional defect center coordinates are integrated with the mold temperature field and pressure time series data recorded during the die casting process, and spatiotemporal alignment is performed to construct five-dimensional spatiotemporal field data; Five-dimensional spatiotemporal field data is input and a reinforcement learning agent model is constructed through a deep Q-network algorithm. Evolutionary action sequences are then explored and generated through physical constraint strategies. The evolutionary action sequence is regularized and inverted using multi-field coupled partial differential equations to reconstruct the evolutionary path of the internal defect region.

7. The intelligent optimization method for aluminum alloy die casting forming process as described in claim 6, characterized in that: The weight parameters of the reverse-engineered graph neural network are corrected and updated. The specific steps are as follows: Based on the evolution path of the reconstructed internal defect region, a holographic defect energy field is constructed in the five-dimensional anti-de Sitter spacetime, with the defect formation time point mapped as time coordinates and the location point mapped as boundary space coordinates; The holographic defect energy field is input into a non-commutative differential operator, and the spatiotemporal invariants are extracted as quantum geometric features through quantum covariant gradient; A pre-trained string theory matrix model is used to perform dimensionality renormalization on quantum geometric features to generate a modified tensor for the weights of the graph neural network. The modified tensor is injected into the backpropagation flow of the graph neural network to perform the update.

8. An intelligent optimization system for aluminum alloy die casting forming process, based on the intelligent optimization method for aluminum alloy die casting forming process according to any one of claims 1 to 7, characterized in that: It includes an element acquisition module, an intelligent compensation module, a linkage execution module, a defect analysis module, a time-tracing inversion module, and a reverse update module; The element acquisition module is used to collect the content of multiple elements in the molten aluminum alloy in real time and combine the content of multiple elements into a three-dimensional material gene vector. The intelligent compensation module is used to input the three-dimensional material gene vector into a pre-trained graph neural network to calculate the injection speed compensation coefficient and the pressurization pressure compensation coefficient. The linkage execution module is used to adjust the real-time injection speed and boosting pressure setting values ​​of the die casting machine based on the injection speed compensation coefficient and the boosting pressure compensation coefficient, and to perform the die casting operation to generate the die casting. The defect analysis module is used to scan the die-cast casting to obtain a three-dimensional tomographic image, identify the internal defect region in the three-dimensional tomographic image through a deep learning segmentation algorithm, and generate four-dimensional defect center coordinates. The time-tracing inversion module is used to align the geometric center coordinates of the defect with the time-series data of mold temperature field and pressure recorded during the die-casting process in time and space. The input is used to construct a reinforcement learning surrogate model through a deep Q-network algorithm to reconstruct the evolution path of the internal defect region. The reverse update module is used to reverse-correct the weight parameters of the graph neural network and update them based on the defect formation time and location points identified in the evolution path.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent optimization method for the aluminum alloy die-casting forming process according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent optimization method for the aluminum alloy die casting process according to any one of claims 1 to 7.

Images